glue_drop and glue_take in bad order

[Vincent-Belliard]

When a shared reference is assign, the glue_drop function is called with the old value before the glue_take function is called with the new_value. With an assignment like "self.o = self.o;" the generated program will core.

Of course, it's a silly example. All other examples I tried make a glue_take before calling the assignment (for example if you call a function with the argument "self.o" which will be then assigned) so you should never have this case in a real program.

Maybe it doesn't need to be fixed as it shouldn't happen with a real program.

[nikomatsakis]

A simpler example:

fn main() {
   let mut x = ~3;
   x = x;
   assert *x == 3;
}

That segfaults for me. This should definitely be fixed!

[graydon]

Hm. At some point we had a test for this because I was quite careful to make this work correctly. Wonder when it broke.

[catamorphism]

Still segfaults as of e16f94f

[brson]

Not 0.6

[catamorphism]

Still segfaults as of 64963d6 . Nominating for milestone 5, production-ready

[graydon]

accepted for production-ready milestone

[emberian]

Still crashing

[klutzy]

Still crashing with 73a28e5

[pnkfelix]

accepted for P-high

[nrc]

Working on this

[nrc]

This no longer reproduces (with NIko's test case). I also tried with a managed pointer because apparently glue take/drop is only used for ref counting now and could not repro that way either. I tried to create a testcase as described in the first comment, but TBH I'm not really sure what was required. I tried a few things nothing crashed. I also examined the llvm ir and the 'problem' is that we never create these drop before take patterns.

Anyone got an idea for how else this bug could be repro'ed (I really don't understand the glue stuff, so I might be missing something obvious)? Otherwise I suggest the bug should be closed.

[thestinger]

I'm not sure if we have a test for this already. Perhaps there's one that was marked as ignored.

[alexcrichton]

With a test case checked in, it looks like this can be closed (unless there's another subtler bug going on here).

[nikomatsakis]

I made a big effort to get this right as part of #3511 -- and totally forgot to write tests!